Estimating the destructiveness of crossover on binary tree representations

In some cases, evolutionary algorithms represent individuals as typical binary trees with n leaves and n-1 internal nodes. When designing a crossover operator for a particular representation and application, it is desirable to quantify the operator's destructiveness in order to estimate its effectiveness at using building blocks. For the case of binary tree representations, we present a novel approach for empirically estimating the destructiveness of any crossover operator by computing and summarizing the distribution of Robinson-Foulds distances from the parent to the entire neighborhood of possible children. We demonstrate the approach by quantifying the destructiveness of a popular tree-based crossover operator as applied to the problem of phylogenetic inferencing. We discuss the benefits and limitations of the destructiveness metric.